Skip to main content
Status:🚀 ACTIVE
Date:📅 unknown
Domain:📦 AI-editorialized journal

SPEC-0016: AI-editorialized journal

  • Status: Accepted
  • Date: 2026-07-11 (Phase 2 revision: 2026-07-12)
  • Capability: journal
  • Source packages: internal/journal (journal.go), internal/store (journal.go, journal_calendar.go, schema.go v11), internal/config (config.go DefaultDigestPrompt), internal/cli (journal.go), internal/web (journal.go, templates/journal.html, templates/home.html)
  • Related ADRs: 📝 ADR-0023 (AI-editorialized journal), 📝 ADR-0011 (contact facts extraction), 📝 ADR-0010 (security & privacy posture)
  • Related specs: SPEC-0005 (contact facts) — the sibling extraction feature

Overview​

msgbrowse builds a day-by-day journal over the local archive in two layers. The mechanical layer is a deterministic per-day rollup (counts and top senders) derived entirely from messages with no LLM and no network egress; it is always built. The digest layer is one prose summary per day written by the configured chat model, cached in SQLite and keyed by (day, model, prompt_version); it is the only network egress and is gated on journal.digest_enabled. Both layers bucket messages by UTC calendar day, honor journal.exclude_conversations before any content is assembled, and are re-ingest-safe (day-keyed, no foreign key to messages).

Phase 2 upgraded the digest from a single prose paragraph to a structured editorial object — the LLM now returns a JSON document (summary, people, themes, mood, timed highlights, standout media, notable links) parsed tolerantly and cached alongside a plain-text summary fallback, with mood denormalized into its own column. The read surface changed to match: the flat, keyset-paginated day list (REQ-0016-010) is superseded by a mood-tinted month-calendar navigator with an editorial day card and headline stat tiles (REQ-0016-014). Phase 2 requirements (REQ-0016-011 through REQ-0016-016) add to the Phase 1 requirements below; none of the two-layer, UTC-bucketing, cache-key, or privacy invariants (REQ-0016-001 through REQ-0016-009) changed.

Requirements​

REQ-0016-001: Mechanical day journal, deterministic and egress-free​

The mechanical journal MUST be a per-day rollup — message count, conversation count, per-source counts, and top senders — derived solely from the local messages table, with NO LLM call and NO network egress. It MUST be built on every msgbrowse journal run regardless of journal.digest_enabled. Rebuilding it MUST be idempotent: a re-run over unchanged messages MUST produce the same journal_days rows (an upsert keyed on day), never duplicates.

Scenario: Mechanical build never egresses​

  • Given an imported archive and journal.digest_enabled = false
  • When the user runs msgbrowse journal
  • Then per-day journal_days rows are written from messages alone, no call is made to llm.base_url, and re-running produces identical rows.

REQ-0016-002: UTC day bucketing​

A day MUST be derived as substr(ts,1,10) (equivalently date(ts_unix,'unixepoch')). Because messages.ts_unix is the wall-clock string parsed as UTC, day bucketing MUST NOT apply any timezone conversion; using 'localtime' (which would double-shift the already-UTC value and misfile messages across midnight) is prohibited. The mechanical rollup and the transcript assembled for a digest MUST use the same UTC day definition.

Scenario: A late-night message stays on its own UTC day​

  • Given a message with ts = '2026-07-01 23:30:00' (ts_unix = that instant as UTC)
  • When the journal is built
  • Then it is counted under day 2026-07-01, and no 'localtime' conversion shifts it into an adjacent day.

REQ-0016-003: Opt-in digest as the single egress​

The LLM digest MUST be produced only when journal.digest_enabled is true, and each digest MUST be the only network egress — one llm.Chat call per day to llm.base_url using llm.chat_model. Import and serve MUST NOT trigger digest generation. A per-day digest call MUST use a longer per-call context timeout (~180s) than the facts default, because a full day's transcript can be large.

Scenario: Digest is explicit and gated​

  • Given journal.digest_enabled = true and days with no cached digest
  • When the user runs msgbrowse journal
  • Then each eligible day is summarized via llm.base_url; running signal-import or serve alone never calls the LLM, and setting digest_enabled = false produces the mechanical journal with zero LLM calls.

REQ-0016-004: Digest cache keyed by day, model, and prompt version​

Each digest MUST be cached in journal_digests (PRIMARY KEY day) alongside the model and prompt_version that produced it, where prompt_version = sha256hex(lower(trim(effective DigestPrompt))) (the same normalization recipe as factHash). A day whose cached (model, prompt_version) matches the current effective values MUST be skipped with no LLM call. Editing journal.digest_prompt or switching llm.chat_model MUST make previously-digested days eligible again on the next run. There MUST be no foreign key from journal_digests (or journal_days) to messages.

Scenario: Prompt change invalidates the cache​

  • Given a day already digested under prompt version P1 and model M
  • When journal.digest_prompt is edited (yielding version P2) and msgbrowse journal runs again
  • Then that day is re-digested and its row is updated to (M, P2); with the prompt and model unchanged instead, the day is skipped with no LLM call.

REQ-0016-005: One digest per day​

The digest MUST be a single cross-thread summary per calendar day (matching config.DefaultDigestPrompt, "summarizing one day"), keyed on day alone. Per-conversation digests are OUT OF SCOPE for this capability.

Scenario: Multi-thread day yields one digest​

  • Given a day with messages across several conversations
  • When the digest for that day is generated
  • Then exactly one journal_digests row is written for that day, summarizing the day as a whole.

REQ-0016-006: Honor the exclude list before assembly​

journal.exclude_conversations MUST be applied during day enumeration and transcript assembly, before any message content is gathered, so an excluded conversation's content never reaches the assembled transcript or the LLM. This MUST NOT affect a day's inclusion when other, non-excluded conversations have messages that day.

Scenario: Excluded thread is never sent​

  • Given a conversation named in journal.exclude_conversations with messages on a given day
  • When that day's digest is generated
  • Then the excluded conversation's content is absent from the transcript sent to llm.base_url, while non-excluded threads from the same day are summarized normally.

REQ-0016-007: CLI flags and run semantics​

msgbrowse journal MUST expose --since YYYY-MM-DD, --backfill, --regenerate, and --dry-run. A default run MUST be incremental: it digests every day whose cached digest is absent or stale by (model, prompt_version). --backfill MUST apply the same eligibility across all history with no day cap. --regenerate MUST wipe all cached digests and rebuild them. --since MUST set a day floor (process only days on or after the given date).

Scenario: Incremental default skips current days​

  • Given some days already digested under the current model and prompt and some not
  • When msgbrowse journal runs with no flags
  • Then only days lacking a current digest are sent to the LLM; already-current days are skipped.

Scenario: Regenerate rebuilds everything​

  • Given a fully-digested history
  • When msgbrowse journal --regenerate runs
  • Then all cached digests are wiped and re-generated for every eligible day.

REQ-0016-008: Per-run day cap with remaining count​

An incremental run MUST cap the number of days it digests at journal.max_days_per_run (0 = unbounded) so a scheduled run has a bounded cost, and MUST report how many eligible days remain after the cap so the operator can schedule follow-up runs. --backfill MUST ignore the cap.

Scenario: Cap bounds a cron run and reports the remainder​

  • Given 50 days eligible for a digest and journal.max_days_per_run = 20
  • When msgbrowse journal runs
  • Then 20 days are digested and the run reports 30 days remaining; a subsequent run digests the next batch.

REQ-0016-009: Dry-run makes zero LLM calls​

--dry-run MUST make NO llm.Chat calls. It MUST enumerate the days that lack a current digest and estimate input tokens with a local len(runes)/4 heuristic. It MUST NOT print a dollar/cost figure — the codebase has no tokenizer or price table and does not decode the provider's usage object, so a real cost estimate is explicitly future work.

Scenario: Dry-run reports days and a token estimate only​

  • Given several days lacking a current digest
  • When msgbrowse journal --dry-run runs
  • Then it prints the eligible-day count and an approximate input-token total (runes/4), makes no network call, and prints no dollar amount.

REQ-0016-010: Journal web page and home quick-link​

Phase 2 (superseded presentation): the flat, keyset-paginated day list this requirement describes has been replaced by the mood-tinted month-calendar navigator specified in REQ-0016-014. The load-bearing invariants below carry forward unchanged — a day renders its cached digest when present and falls back to the mechanical rollup otherwise, an empty archive renders an empty state, the home page offers a fourth quick-link tile to /journal, and rendering the page issues no LLM call. Only the day-navigation surface changed (keyset list → calendar grid + editorial day card); the scenarios below record the Phase 1 behavior.

GET /journal MUST list days newest-first, showing each day's digest when one is cached and falling back to the mechanical rollup when it is not, with keyset pagination over days and an empty state when no days exist. The home page MUST offer a fourth quick-link icon tile linking to /journal. Rendering the page MUST NOT trigger any LLM call.

Scenario: Days render newest-first with fallback​

  • Given a day with a cached digest and an earlier day with only a mechanical rollup
  • When the user opens /journal
  • Then the digested day appears first showing its prose, the earlier day shows its counts and top senders, older days page in via keyset pagination, and no LLM call is made.

Scenario: Empty state​

  • Given an archive with no imported messages
  • When the user opens /journal
  • Then the page renders an empty state rather than an error or a blank list.

REQ-0016-011: Structured editorial digest with tolerant parsing​

The digest MUST be produced as a single JSON object with the keys summary (string), people (string array), themes (string array), mood (enum string), highlights (array of {text, time} objects), standout_media (string array), and notable_links (string array), matching config.DefaultDigestPrompt. The response MUST be parsed tolerantly: the object is extracted from the first { to the last } (tolerating markdown fences or surrounding prose), and every field is coerced rather than rejected — each list item is trimmed and empty items are dropped, an unknown mood is coerced to "neutral", and a malformed HH:MM highlight time is blanked ("") while its text is kept. A response that carries no JSON object or whose summary is empty MUST be treated exactly like an empty body: the day is skipped-and-logged (counted as skipped, no row written) so a re-run retries it, and it MUST NOT wedge the resumable run. The validated digest MUST be re-canonicalized to JSON before it is stored.

Scenario: Fenced, partly-malformed response is coerced, not rejected​

  • Given an LLM response wrapping the JSON object in markdown fences with a blank people entry, an unknown mood, and one highlight whose time is not HH:MM
  • When the digest is parsed
  • Then the object is extracted from the fences, the empty people item is dropped, mood becomes "neutral", the bad highlight time is blanked while its text is retained, and the canonicalized digest is stored.

Scenario: No-JSON or empty-summary response is skipped, not persisted​

  • Given an LLM response containing no JSON object (or a JSON object with an empty summary)
  • When that day is digested
  • Then no journal_digests row is written, the day is logged and counted as skipped, and a subsequent run treats the day as still eligible.

REQ-0016-012: Mood is a fixed enum, coerced and denormalized​

mood MUST be one of a fixed allowlist — journal.Moods = upbeat, neutral, quiet, tense — kept in sync with config.DefaultDigestPrompt. Any value outside the allowlist (including an empty or unrecognized string) MUST be coerced to "neutral" at parse time. The coerced mood MUST be denormalized into the journal_digests.mood column so the calendar and stat reads can tint day cells without unmarshaling the structured blob per day.

Scenario: Unknown mood degrades to neutral​

  • Given a model that returns "mood": "chaotic"
  • When the digest is parsed and stored
  • Then journal_digests.mood is written as "neutral", and the calendar reads the column directly without parsing structured.

REQ-0016-013: Structured columns via in-place schema amendment and prompt-version re-derivation​

journal_digests MUST carry a structured column (the canonical JSON of the validated digest) and a mood column, both defaulting to '' so a prose-only or legacy row reads cleanly; body MUST remain the plain-text summary used as the fallback and empty-response guard. Because these columns were added by amending schemaV11 in place (safe only while v11 is unmerged/unshipped), and because journal_digests is a derived cache with no source-of-truth data, no schema version bump and no --regenerate are required to populate them: editing config.DefaultDigestPrompt changes prompt_version, which makes every existing prose-era digest eligible again and re-derives it in structured form on the next run — a free data migration. A database already at user_version = 11 from before the amendment MUST be understood NOT to retroactively gain the new columns (the CREATE TABLE IF NOT EXISTS is a no-op on an existing table); the documented remedy is to recreate the derived-cache database (it is rebuildable from messages), not to hand-alter the table.

Scenario: Prompt-version bump re-derives prose rows as structured​

  • Given days digested before Phase 2 (prose body, empty structured/mood) and a DefaultDigestPrompt edited to demand the JSON object
  • When msgbrowse journal runs against a database that has the amended journal_digests shape
  • Then each affected day is re-digested under the new prompt_version, and its row is updated in place with canonical structured JSON and a denormalized mood, with no schema bump and no --regenerate.

REQ-0016-014: Mood-tinted month-calendar navigator with editorial day card and stat tiles​

GET /journal MUST render a mood-tinted month calendar navigator: year tabs (newest first), a fixed month grid whose present days are tinted by that day's mood and annotated with a message-count subscript, previous/next-month controls, and a mood legend. It MUST render an editorial day card for the selected day (summary, timed highlights, people, themes, standout media, notable links) and a row of stat tiles (days-with-entries, longest streak, most-active weekday, peak hour). Navigation MUST be by boosted query parameters (?year, ?month, ?day) with no client-side state. The bare /journal MUST open on the newest day's card; a year tab (?year with no month) MUST open on that year's latest active month, never an empty January. When the selected day has no structured digest, the card MUST fall back to the prose body, then to the mechanical top-senders. Rendering MUST issue no LLM call, and an archive with no journal MUST render the empty state.

Scenario: Bare landing opens the newest day​

  • Given a built journal whose newest day is 2026-07-11
  • When the user opens /journal with no query parameters
  • Then the calendar shows July 2026, 2026-07-11 is selected and its editorial card is rendered, and no LLM call is made.

Scenario: Year tab opens the year's latest active month​

  • Given a year whose activity begins in March
  • When the user clicks that ?year tab (no month specified)
  • Then the calendar opens on that year's most recent active month, not January, and no empty grid is shown.

Scenario: Day cell tinted by mood with a count subscript​

  • Given a day with a cached digest whose mood is tense and 42 messages
  • When the month grid renders
  • Then that day's cell carries the tense mood tint and a 42 count subscript and links to ?day= for that date; days without content are inert blanks.

REQ-0016-015: Calendar read surface, UTC-bucketed and exclude-honoring​

The calendar/stat reads MUST be served by dedicated store methods: JournalMonth (≤31 rows off journal_days joined to journal_digests.mood), JournalStats, GetJournalDay (rollup joined with its digest, structured and mood included), LatestJournalDay and LatestJournalDayInYear, and JournalYears. JournalStats MUST derive days-with-entries and longest streak from the journal_days key set in Go (longestStreak, adjacency by date arithmetic so month/year rollovers count), and most-active weekday and peak hour as argmax GROUP BY reads over messages (journalArgmax), a year bounded by a sargable ts_unix range (not a date(ts_unix) wrap). All day bucketing MUST be UTC. Year 0 means all-time (a full scan), but the web layer MUST always pass a concrete year. JournalStats MUST honor journal.exclude_conversations — the same denylist journal_days was built with — so excluded threads never inflate the stat tiles.

Scenario: Streak counts a month rollover​

  • Given journal days 2026-01-30, 2026-01-31, 2026-02-01
  • When JournalStats computes the longest streak
  • Then the streak is 3, because adjacency is date arithmetic, not string succession.

Scenario: Stats honor the exclude denylist​

  • Given journal.exclude_conversations naming a high-volume thread
  • When the stat tiles are computed for a year
  • Then the excluded thread's messages are absent from the weekday/peak-hour argmax and from the day/streak counts, matching the journal the day rollups were built with.

REQ-0016-016: Untrusted structured fields are escaped; no attacker-controlled markup​

Every structured-digest field is model-derived and MUST be treated as untrusted output. All rendered fields (summary, highlights, people, themes, standout media, notable links) MUST be emitted through html/template auto-escaping. Notable links MUST render as text only, never as a raw href. Mood tints MUST be applied as CSS classes keyed by the fixed mood enum (cal-day--<mood>), never an attacker-supplied class or inline style. The page MUST remain compatible with the site CSP (style-src 'self', no 'unsafe-inline'): no style= attribute may carry model-derived values.

  • Given a digest whose summary contains <script> and whose notable_links contains a javascript: URL
  • When the day card renders
  • Then the <script> is HTML-escaped as inert text, the link is shown as escaped text with no clickable href, no inline style= is emitted, and the mood tint is a fixed cal-day--<mood> class.